Techniques for adaptive machine learning and their applications to recommendation systems
Event details
| Date | 24.06.2016 |
| Hour | 09:00 › 11:00 |
| Speaker | Fei Mi |
| Location | |
| Category | Conferences - Seminars |
EDIC Candidacy Exam
Exam President: Prof. Karl Aberer
Thesis Director: Prof. Boi Faltings
Co-examiner: Prof. Volkan Cevher
Background papers:
Collaborative filtering with temporal dynamics, by Y. Koren
Bayesian variable order Markov models, by C. Dimitrakakis
A contextual-bandit approach to personalized news article recommendation, by Lihong Li et al.
Abstract
Machine learning is often used to acquire knowledge in domains that undergo frequent changes, such as networks, social media, or markets. These frequent changes pose a chal- lenge to most machine learning methods as they have difficulty adapting. In this proposal, we discuss three existing works of modeling temporal factors which shed lights on our proposal for building adaptive machine learning models. The first work explicitly models temporal factors for latent factor models and neighborhood models. The later two works formalize the problem as a sequence prediction problem and propose two online algorithms based on context tree structure and multi-arm bandit formulation respectively. To start with our research, we compare different recommendation methods, including a new context tree (CT) method. The results show that the CT recommender performs better than other baseline methods due to its adaptation to changes in the domain, which highlights the importance of considering this aspect when using machine learning.
Exam President: Prof. Karl Aberer
Thesis Director: Prof. Boi Faltings
Co-examiner: Prof. Volkan Cevher
Background papers:
Collaborative filtering with temporal dynamics, by Y. Koren
Bayesian variable order Markov models, by C. Dimitrakakis
A contextual-bandit approach to personalized news article recommendation, by Lihong Li et al.
Abstract
Machine learning is often used to acquire knowledge in domains that undergo frequent changes, such as networks, social media, or markets. These frequent changes pose a chal- lenge to most machine learning methods as they have difficulty adapting. In this proposal, we discuss three existing works of modeling temporal factors which shed lights on our proposal for building adaptive machine learning models. The first work explicitly models temporal factors for latent factor models and neighborhood models. The later two works formalize the problem as a sequence prediction problem and propose two online algorithms based on context tree structure and multi-arm bandit formulation respectively. To start with our research, we compare different recommendation methods, including a new context tree (CT) method. The results show that the CT recommender performs better than other baseline methods due to its adaptation to changes in the domain, which highlights the importance of considering this aspect when using machine learning.
Practical information
- General public
- Free
Contact
- Cecilia Chapuis EDIC